rl-for-topic-models

Reinforcement Learning for Topic Models

Code accompanying the paper Reinforcement Learning for Topic Models in Findings of the Association for Computational Linguistics: ACL 2023.

Table of Contents

• Install
• Training
• Data
• Experiments
• Evaluations

INSTALL

conda create -n rl-for-topic-models python=3.9
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

TRAINING

• update the configs in model/decoder_network.py and trainer/config.py with your training settings.
• run python train.py /path/to/data/pickle
  • add --test if your data has a test subset

DATA

Pre-training is a Hot Topic

Tweets 2011

• get raw data from https://trec.nist.gov/data/tweets
• put processed Tweet.txt in the data/raw/texts folder
• run python data/dataset.py tweets2011

Stack Overflow

• get data from https://github.com/qiang2100/STTM
• put StackOverflow.txt in the data/raw/texts folder
• run python data/dataset.py stackoverflow

Google News

• get data from https://github.com/qiang2100/STTM
• put GoogleNews.txt in the data/raw/texts folder
• run python data/dataset.py googlenews

Wiki 20k

• get data from https://github.com/vinid/data
• put dbpedia_sample_abstract_20k_unprep.txt in the data/raw/texts folder
• run python data/dataset.py wiki20k

20 Newsgroups

• run python data/dataset.py 20ng

Other Comparison Papers

New York Times

• get raw data from https://catalog.ldc.upenn.edu/LDC2008T19
• put tarball in data/raw/nyt
• run pip install beautifulsoup4 lxml
• run python data/raw/nyt/nyt_untar.py
• run python data/dataset.py nytcorpus

Contrastive Learning for Neural Topic Model

• get raw data from https://github.com/nguyentthong/CLNTM
• put zipfile in data/raw/clntm
• run python data/raw/clntm/scholar_unzip.py
• run python data/dataset.py contrastive

Benchmarking Neural Topic Models

• get data from https://github.com/smutahoang/ntm
• put *_preprocessed_data.txt in data/raw/texts

Our Experiments

20 Newsgroups

• run python data/dataset.py 20ng --mwl 3

Custom Data

• run python data/dataset.py custom --train_file /path/to/train/file --save_name /path/to/save/name
  • other arguments can be found at the bottom of data/dataset.py

EXPERIMENTS

• update search_dict in run_experiments.py with your hyperparameter search values.
• run python run_experiments.py experiment_name num_seeds --meta_seed meta_seed
  • meta seed will be randomly chosen from random.randint(0, 2 ** 32) if not included as argument

EVALUATIONS

Compute Metrics

• run python -m evals.compute_metrics topk num_experiments num_seeds /path/to/data/pickle /path/to/experiment

Dataset Statistics

• run python -m evals.dataset_stats /path/to/data/pickle

Dropout Sweep

• run experiment with hyperparameters from paper
• run python -m evals.dropout_sweep /path/to/dropout/experiment

Benchmarking Neural Topic Models Plot

• run empirical_studies/examine_models.py from https://github.com/smutahoang/ntm with top_ks = [10]
• move the run.*.pkl files into the ntm_runs folder
• run experiments with hyperparameters from paper
  • call them ntm_20news_sweep, ntm_snippets_sweep, ntm_w2e_sweep, and ntm_w2e_text_sweep
• run python -m evals.plot_ntm_results

Saving Topic Words

• run python -m evals.save_topic_words topk /path/to/data/pickle /path/to/experiment/experiment_num/seeds/seed_num/seed_num_plotting_arrays.pkl